WFGY/ProblemMap/Atlas/atlas-to-ai-adapter-v1.md
2026-03-19 17:36:37 +08:00

20 KiB

Atlas-to-AI Adapter v1 🤖

Official AI-facing routing and diagnosis layer for Atlas v1

Quick links:


If the Atlas core defines the structure and the casebook teaches that structure through examples, this file is the layer that makes the Atlas usable inside real AI routing, diagnosis, and debugging flows. 🧭

This document does not create a new Atlas.

It takes the frozen Atlas grammar and turns it into a disciplined operational layer for model-facing use.

Short version:

the Atlas defines the map
the casebook teaches the map
the adapter lets AI systems operate on the map

That is the job of this file.


Quick start 🚀

I am new to the adapter layer

Use this path:

  1. read Atlas Final Freeze v1
  2. read Canonical Casebook v1
  3. read this file
  4. read Adapter Runtime Modes v1
  5. read Adapter Failure Discipline v1

I already know the Atlas and want the shortest route

Start here:

  1. read Section 5 for the Core Contract
  2. read Section 6 for the Routing Prompt Stack
  3. read Section 8 and Section 9 for the runtime mode system
  4. read Section 12 for adapter failure discipline
  5. read Section 16 for patch interface

Shortest possible reading:

route first
explain the cut
state the broken invariant
keep confidence honest
preview repair without pretending full closure


0. Document status 🚦

This document is the frozen first formal AI-facing adapter for Atlas v1.

Its purpose is to turn Atlas logic into a reusable operational layer for:

  • AI-assisted routing
  • debugging support
  • issue triage
  • workflow diagnosis
  • benchmark failure classification
  • teaching and onboarding
  • compact model-facing routing

This document is frozen as Atlas-to-AI Adapter v1.

It is frozen not because all future adapter work is complete, but because the first stable adapter body is now strong enough to support real use, teaching, reuse, and patch-driven growth.

Future work should proceed through adapter patch mode, not through silent rewriting of the adapter contract.


1. What this adapter is 🧭

The Atlas defines the structural grammar.

The casebook teaches how that grammar behaves through examples.

The adapter turns that grammar into a model-facing routing layer.

In short:

the Atlas defines the map
the casebook teaches the map
the adapter lets AI systems operate on the map

This adapter exists to help a model do five things more reliably:

  1. route a case into the correct family
  2. distinguish primary from secondary family pressure
  3. state the broken invariant explicitly
  4. point to the best current fit
  5. give the correct first repair direction without pretending full root-cause closure

2. What Atlas-to-AI Adapter v1 contains 🧱

Atlas-to-AI Adapter v1 contains five major layers.

A. Core Contract v1

This defines:

  • minimum input expectations
  • minimum output structure
  • mandatory routing fields
  • confidence discipline
  • evidence discipline

B. Routing Prompt Stack v1

This defines:

  • the order of reasoning constraints
  • the role of Atlas rules
  • the role of boundary rules
  • the role of fit discipline
  • the role of output discipline

C. Casebook Injection Layer v1

This defines:

  • when to use family anchors
  • when to use boundary exemplars
  • when to use repair exemplars
  • how many exemplars to use
  • how exemplars should support, but not replace, structural routing

D. Runtime Mode Pack v1

This defines four operational modes:

  • Strict Routing Mode
  • Teaching Mode
  • Repair-First Preview Mode
  • Compact Routing Mode

E. Patch Interface and Failure Discipline v1

This defines:

  • how future adapter updates should happen
  • what counts as small, medium, or large patch
  • what the adapter must never do, even when compressed or optimized

3. What this adapter claims

Atlas-to-AI Adapter v1 claims that the following are now stable enough to freeze:

  • the first adapter contract
  • the first routing prompt stack
  • the first casebook injection rules
  • the first runtime mode system
  • the first adapter patch interface
  • the first formal adapter failure discipline

This means the adapter is now stable enough to support:

  • AI routing prompts
  • AI-assisted troubleshooting
  • model-facing routing demos
  • onboarding and teaching
  • product-facing AI flows
  • TXT pack construction
  • future adapter patch waves

4. What this adapter does not claim 🔍

Atlas-to-AI Adapter v1 does not claim that:

  • every model will route perfectly under all conditions
  • all prompts, platforms, and contexts behave identically
  • all future compact forms are already optimized
  • the adapter already contains the full repair system
  • the adapter replaces the Atlas core
  • the adapter replaces the casebook
  • no future adapter patching is needed

Atlas-to-AI Adapter v1 claims only that:

the first stable AI-facing routing and diagnosis layer of the Atlas now exists and is strong enough to freeze as a formal reusable version


5. Core Contract v1 📜

The core contract is the minimum discipline that keeps the adapter stable.

5.1 Minimum input expectation

The adapter expects some form of failure-bearing input, such as:

  • an issue report
  • a failing output
  • a benchmark failure case
  • a workflow collapse description
  • a short debugging scenario
  • a casebook-like teaching case
  • a structured internal failure note

The adapter does not require perfect completeness.

But it does require enough content to justify a routing attempt.

5.2 Minimum output expectation

A valid adapter output must preserve routing structure.

At minimum, the adapter should be able to produce:

  • primary_family
  • secondary_family
  • why_primary_not_secondary
  • broken_invariant
  • best_current_fit
  • fit_level
  • fix_surface_direction
  • confidence
  • evidence_sufficiency

Where ambiguity or no-fit is materially relevant, those fields must also be made explicit.

5.3 Contract principle

The adapter is not allowed to skip structural justification in order to sound smoother, more confident, or more fluent.

Correct routing discipline is more important than rhetorical polish.


6. Routing Prompt Stack v1 🧠

The routing stack defines how the adapter should think in ordered layers.

Layer 1 · Core routing objective

First determine:

  • what kind of failure this is
  • which family is primary
  • which neighboring family exerts secondary pressure

Layer 2 · Boundary discipline

Before final routing, check the major cuts:

  • F1 vs F7
  • F5 vs F6
  • F3 vs F4
  • F2 vs F7
  • F4 vs F5
  • F1 vs F5

If the case sits near a known high-pressure cut, this must be reflected in the output.

Layer 3 · Broken invariant discipline

The adapter must state what invariant is broken first.

This matters because family assignment without invariant language becomes too soft and too rhetorical.

Layer 4 · Best-fit discipline

The adapter should map the case into:

  • family
  • node or family-entry
  • or edge-fit wording when the case sits at a strong boundary

The adapter should not pretend full subtree closure if the fit is still local, edge-based, or patch-sensitive.

Layer 5 · Repair-facing discipline

Only after routing should the adapter provide first repair direction.

Repair output must remain:

  • routing-sensitive
  • first-move oriented
  • non-grandiose
  • honest about uncertainty

7. Casebook Injection Layer v1 📚

The adapter is allowed to use canonical exemplars, but only under discipline.

7.1 Purpose of exemplar injection

Exemplars exist to support:

  • family teaching
  • boundary teaching
  • repair-preview teaching

They do not exist to replace structural routing rules.

7.2 Injection hierarchy

Pack A · Family Anchor Pack

Use when the goal is to help stabilize a first family interpretation.

Pack B · Boundary Teaching Pack

Use when the case sits near a major family cut and needs stronger why-not-neighbor reasoning.

Pack C · Repair Teaching Pack

Use when the case is already routed and the system needs a more stable first repair preview.

7.3 Injection rule

Exemplars may support routing, but they must never override:

  • frozen family rules
  • frozen boundary rules
  • broken-invariant discipline
  • evidence discipline

7.4 Minimal injection principle

Use as little exemplar support as needed.

The adapter should not become exemplar-heavy by default.

It should remain rule-first, exemplar-supported.


8. Runtime Mode Pack v1 ⚙️

Atlas-to-AI Adapter v1 defines four formal runtime modes.


Mode 1 · Strict Routing Mode

Purpose
Stable classification and routing.

Best for

  • issue triage
  • benchmark fail routing
  • workflow failure routing
  • real debugging pipelines
  • batch case routing

Behavior

  • family routing first
  • explicit why-primary-not-secondary
  • explicit broken invariant
  • explicit fit level
  • fix direction restricted to first move
  • no unjustified high confidence under weak evidence

Style

Short, hard, structured, low-noise.

Risk posture

This is the safest and least hallucination-prone default mode.


Mode 2 · Teaching Mode

Purpose
Teach users or AI systems how and why a case is being cut in a particular way.

Best for

  • onboarding
  • documentation support
  • learning sessions
  • collaborator training
  • boundary explanation

Behavior

  • preserves full structural output
  • explains why-not-neighbor more fully
  • may name the relevant boundary rule
  • may explain what goes wrong if the case is cut incorrectly

Style

More explanatory than Strict Mode, but still structure-first.

Risk posture

Teaching must not degrade into freeform essay behavior.

The adapter still needs to respect the contract.


Mode 3 · Repair-First Preview Mode

Purpose
Give a useful first repair move after correct routing.

Best for

  • demos
  • product UX
  • assistant workflows
  • route-and-suggest systems
  • first-step debug support

Behavior

  • route first
  • then give 1 to 3 first repair directions
  • may mention misrepair risk
  • does not provide a full repair screenplay

Style

Structured, slightly more action-oriented.

Risk posture

This mode must not overpromise.

It gives the first move, not full closure.


Mode 4 · Compact Routing Mode

Purpose
Support short context, low-token, or embedded routing situations.

Best for

  • short-window models
  • batch routing
  • embedded pipeline use
  • lightweight agent layers
  • low-cost deployment

Behavior

  • outputs the shortest viable structure
  • keeps a compressed why-primary-not-secondary line
  • preserves fit level
  • preserves confidence
  • preserves evidence sufficiency
  • keeps repair direction to 1 or 2 items at most

Style

Short, hard, token-conscious.

Risk posture

Compact mode is allowed to compress, but not to erase structural essentials.


9. Runtime Mode Matrix v1 📊

Mode Core purpose Exemplar use Explanation depth Repair strength Best use case
Strict Routing stable classification low medium low real routing and triage
Teaching teaching cuts medium high low to medium onboarding and teaching
Repair-First Preview route then first repair move medium to high medium medium demos and engineering helpers
Compact Routing low-token fast routing very low low very low short-window and batch use

This matrix is frozen because it provides a clean first operational grammar for deployment.


10. Minimum Output Schema 🧩

The adapter must preserve different minimum requirements by mode.

10.1 Strict Routing minimum output

Must include:

  • primary_family
  • secondary_family
  • why_primary_not_secondary
  • broken_invariant
  • best_current_fit
  • fit_level
  • ambiguous
  • no_fit
  • fix_surface_direction
  • confidence
  • evidence_sufficiency

10.2 Teaching Mode minimum output

Must include everything in Strict Mode, plus:

  • boundary_rule_explanation
  • misroute_risk_note

10.3 Repair-First Preview minimum output

Must include everything in Strict Mode, plus:

  • clearer first repair move note
  • misrepair risk note

10.4 Compact Routing minimum output

Must include:

  • primary_family
  • secondary_family
  • why_primary_not_secondary
  • broken_invariant
  • best_current_fit
  • fit_level
  • fix_surface_direction
  • confidence
  • evidence_sufficiency

Compact mode may omit ambiguity and no-fit explanation only when both are false.

If ambiguity or no_fit is true, the reason must still be stated.


11. Confidence and Evidence Discipline 🔒

The adapter must obey two permanent disciplines.

11.1 Confidence discipline

The adapter must not use high confidence merely because a case sounds familiar, famous, or rhetorically strong.

Confidence must track actual structural evidence.

11.2 Evidence sufficiency discipline

If the input evidence is weak, sparse, or highly compressed, the adapter must say so.

Weak evidence should not be hidden behind fluent structural language.

These two disciplines matter because routing grammar collapses if decisiveness becomes disconnected from evidence.


12. Adapter Failure Discipline v1 🛡️

The adapter must obey the following failure discipline.

  1. route first, then repair
  2. primary must correspond to the broken invariant
  3. secondary must reflect real boundary pressure
  4. ambiguity must not be opened merely because a case is difficult
  5. no-fit must not be opened merely because a model is lazy
  6. weak evidence must not produce high confidence
  7. overlays must not silently become primary families
  8. exemplars may support, but not override, structural rules
  9. repair output must remain first-move level, not full-script theater
  10. patches must be versioned, not silently rewritten

These ten rules are the safety rails of the adapter.

For the full discipline layer, see:


13. Relationship to Atlas Final Freeze v1 🔗

This adapter is not the Atlas core.

It depends on the Atlas core.

Use Atlas Final Freeze v1 to determine:

  • family structure
  • routing grammar
  • canonical node logic
  • subtree structure
  • relation logic

Use this adapter to make that structure operational for AI systems.

If the Atlas changes through patch logic, the adapter must follow the Atlas.

The adapter must never quietly redefine the Atlas.


14. Relationship to Canonical Casebook v1 📘

This adapter also depends on the casebook.

Use Canonical Casebook v1 to supply:

  • family anchor exemplars
  • boundary teaching exemplars
  • repair teaching exemplars

The casebook stabilizes the adapter by providing reusable teaching cases.

The adapter operationalizes the casebook by deciding when and how those cases should be injected.

The two are complementary, not interchangeable.


15. Relationship to the Repair Layer 🛠️

Atlas-to-AI Adapter v1 does not contain the full repair system.

What it does contain is the first repair-facing interface.

That means:

  • it can point to the correct repair family
  • it can preview the first move
  • it can warn about misrepair risk
  • it can preserve route-first discipline

What it does not yet do in complete form is provide a full deep repair architecture for every branch.

For deeper repair-facing work, this adapter should later connect to:


16. Patch Interface v1 🔄

The adapter is frozen, but not closed.

Small Patch

Use for:

  • wording refinement
  • confidence tuning
  • fit guidance improvement
  • fix-direction wording adjustments
  • exemplar injection default improvements

Medium Patch

Use for:

  • subtree-aware routing upgrades
  • new high-pressure mini-stacks
  • runtime behavior tuning
  • compact routing optimization
  • selector logic expansion

Large Patch

Only use if:

  • the Atlas family structure changes
  • major family cuts collapse repeatedly
  • exemplar logic becomes systemically unstable
  • runtime mode architecture must be redesigned

Current status

No large-patch pressure is currently justified.


17. What this adapter enables 🚀

Atlas-to-AI Adapter v1 now makes the following possible:

  • AI-assisted family routing
  • compact model-facing classification
  • teaching-oriented routing sessions
  • route-first repair preview
  • structured troubleshooting prompts
  • TXT pack construction
  • adapter patch waves
  • AI reuse without flattening the Atlas core

This is the first point where the Atlas can be treated not only as a concept map, but as an operational AI-facing layer.


18. Official freeze statement 📣

The correct formal statement is:

Atlas-to-AI Adapter v1 is complete as the first formal AI-facing routing and diagnosis layer of the Atlas system.
The adapter contract, routing stack, casebook injection layer, runtime modes, patch interface, and adapter failure discipline are now stable enough for formal reuse.
Future adapter work should proceed in patch mode.


Next steps

After this page, most readers continue with:

  1. Open Adapter Runtime Modes v1
  2. Open Adapter Failure Discipline v1
  3. Open Canonical Casebook v1
  4. Open Fixes Hub

If you want the broader Atlas surface:


19. One-line version 🌍

Atlas-to-AI Adapter v1 is the first frozen AI-facing routing and diagnosis layer for Problem Map 3.0 Troubleshooting Atlas.


20. Closing note

A map is powerful.

A teaching system is more powerful.

A routing layer that lets AI systems use that map under discipline is even more powerful.

That is what this adapter is for.

It does not replace the Atlas.

It makes the Atlas operational.